Bethard cikm2010

Citation

• Title : Who Should I Cite? Learning Literature Search Models from Citation Behavior
• Authors : S. Bethard and D. Jurafsky
• Venue : CIKM 2010

Summary

This paper describes a retrieval model to search relevant existing work in a collection of scientific articles. The authors claim that the model is useful when a researcher wants to conduct a new research outside his/her area of expertise and needs to get familiar with prior work in the field. The model incorporates various text and meta features and uses the citation networks to learn weights of these features.

Dataset

ACL Anthology (~11,000 papers)

Model

Documents are ranked based on their scores.
For scoring, they used linear model between a query Q (project idea) and a document D (existing scientific article) as follows :

$score(Q,D)=\sum _{i}w_{i}\times f_{i}(Q,D)$

Features

• Terms :
  • TF-IDF between Q and D
• Citations :
  • Number of papers that cited D
  • Number of citations for articles in the venue in which D was published
  • Number of citations author of D has received (if there are multiple authors, use one with the most citation counts)
    • A variant using h-index instead of raw citation counts was also explored. An author with h-index h has published h papers each of which has been cited at least h times.
  • PageRank score of document collection calculated over the citation network (instead of the hyperlink network)
• Recency:
  • current year - year D. (intuition : older papers get less scores).
• Similar Topics (100 topics from LDA)
  • Cosine similarity between Q and D
  • Cosine similarity between Q and averaged topic distributions of all other documents which cited D
  • Topic citation counts
  • Entropy of D's topic distribution
  • Entropy of documents-citing-D's mean topic distributions
• Social habits
  • TF-IDF between authors of Q and authors of D (note that this just is a TF-IDF between author lists)
  • TF-IDF between authors of Q and authors of
  • Authors of previously cited articles by authors of Q

Experiments

Feature scores (citation counts, etc.) were log-transformed and scaled to between 0 and 1.
Experimented with two classifiers :

• Logistic regression
• SVM-MAP

Training / Dev / Test split

Results

• Mean average precision on the dev set for different classifiers (using all features)
  

Logistic 50/50 model downsampled the number of negative examples to be the same as the number of positive examples.

• Mean average precision on the dev and test sets using different feature sets (model : SVM-MAP)
  

Feature analysis

Below are weights of each feature used in the experiments